Abstract: This paper introduces global shape modeling by means ofMarkov random fields and describes its use in medicalimage segmentation. The key point positionsrepresenting the shape of an object are assumed to bemultivariate Gaussian distributed with a certaincovariance structure which relates to the Markovproperty with respect to some neighborhood system.Since the neighborhood of a key point potentiallycontains both nearby and long distant key points,global key point interaction is not only realized bypropagated local key point interaction, but alsodirectly by long distant key point interaction. Werestrict ourselves to the subclass of decomposablemodels, since a closed form expression for the maximumlikelihood estimate of the covariance matrix from a setof training shapes is available in this case. Theneighborhood system is either a priori defined orestimated. Our model building procedure is demonstratedfor the 2D shape of spinal vertebra. The suitability ofthe derived shape models is investigated by generatingnew shape samples according to the models. Finding theobject's boundary in a grey value image is formulatedas maximum a posteriori estimation incorporating theshape model as a priori model. Our model-basedsegmentation procedure includes an easy and effectiveinteractive improvement of the segmentation outcome.!17
展开▼
